ABSSNet: Attention-Based Spatial Segmentation Network for Traffic Scene Understanding.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Accurate detection of road and lane lines is critical for autonomous and auxiliary driving systems.
• Complex traffic environments pose significant challenges to existing deep learning models, leading to frequent false and missing detections.
• Conventional Convolutional Neural Networks (CNNs) often fail to fully leverage spatial information within layers.

Purpose of the Study:

• To propose an attention-based spatial segmentation network for enhanced traffic scene understanding.
• To improve the accuracy and reliability of road and lane line detection in challenging real-world scenarios.
• To address the limitations of existing CNNs in utilizing spatial information for precise localization.

Main Methods:

• Developed an attention-based spatial segmentation network incorporating a convolutional attention module.
• Implemented Spatial CNN (SCNN) to enhance information flow within convolutional layers and model spatial relationships.
• Introduced the NWPU Road Dataset, a pixel-level road segmentation dataset, to facilitate research.

Main Results:

• The proposed network effectively improves the utilization of spatial information by neural networks.
• Experimental results demonstrate a significant enhancement in traffic scene understanding capabilities.
• The method leads to a notable reduction in false and missing detections for road and lane lines.

Conclusions:

• The attention-based spatial segmentation network offers a promising solution for accurate road and lane line detection.
• Improved spatial information modeling is key to advancing the reliability of autonomous driving systems.
• The developed NWPU Road Dataset serves as a valuable resource for future research in traffic scene analysis.